10 Waymo Perception Engineer (New Grad) Interview Questions (2026)
Waymo's Perception Engineer new-grad loop in 2026 is a recruiter screen, an OA, and a 5-round virtual onsite with significant emphasis on computer vision, sensor fusion, and the ML pipelines that turn LIDAR and camera data into tracked objects. Perception is the entry point of the autonomy stack — bugs here cascade everywhere — so the bar is calibrated higher than generalist SWE.
By Alex Chen, Founder, InterviewChamp.AI · Last verified
Loop overview
Perception-track new-grads see an extended loop. Recruiter → OA → phone screen → onsite with two coding rounds, one perception-domain round (sensor fusion, camera-LIDAR projection, tracking), one ML round (model architectures, training, evaluation), and one behavioral. C++ is required; Python for ML pipelines. Most candidates have at least a Master's in CS/EE with a vision/robotics specialization.
Behavioral (3)
Why perception specifically (vs planning or controls)?
Frequently askedOutline
Tie to a specific interest: computer vision, sensor processing, ML in safety-critical settings. If you have a vision research paper, project, or course, mention it. Show you understand that perception bugs cascade — your output is everyone else's input.
Tell me about a time you had to balance latency vs accuracy.
Frequently askedOutline
STAR. Perception runs on a tight latency budget — 100ms is typical for a frame. Pick a project (class, research, internship) where you made a similar tradeoff. Discuss the methodology: profile, identify bottlenecks, simplify, evaluate impact on the downstream metric.
Tell me about a research paper or project that influenced how you think about perception.
Occasionally askedOutline
Pick a specific paper. PointNet, Faster R-CNN, CenterPoint, or DETR are all reasonable starting points. Walk through what the paper does, why it mattered, what you took from it. Show you read primary sources, not just summaries.
Technical (3)
Implement a function that, given a list of 3D bounding boxes from two consecutive frames, matches them to produce tracks.
Frequently askedOutline
Hungarian algorithm for optimal assignment. Cost matrix = pairwise IoU or distance. Threshold to reject unlikely matches. Walk through the assignment step. O(N^3) for Hungarian. Discuss alternatives: greedy nearest-neighbor, JPDA for probabilistic tracking. Handle birth (new track) and death (lost track).
Given a 2D point cloud, cluster the points using DBSCAN.
Frequently askedOutline
DBSCAN parameters: epsilon (neighborhood radius), min_samples (core point threshold). For each unvisited point: find neighbors within epsilon; if >= min_samples, start a new cluster and expand; else mark as noise. Walk through expansion (BFS over the neighborhood graph). Discuss the complexity (O(N log N) with a spatial index, O(N^2) without).
Implement non-maximum suppression for bounding boxes.
Frequently askedOutline
Sort detections by confidence descending. For each box, accept it; suppress all subsequent boxes with IoU > threshold. Walk through IoU calculation. O(N^2) naive; O(N log N) with sorted IoU index. Discuss soft-NMS as a less aggressive alternative for crowded scenes.
Domain knowledge (4)
Walk through how you would fuse a LIDAR point cloud with a camera image to produce a colored 3D point cloud.
Frequently askedOutline
Calibration first: extrinsics (rigid transform between LIDAR and camera frames) + camera intrinsics (focal length, principal point, distortion). For each LIDAR point: transform to camera frame, project to image plane (perspective division), look up the pixel color. Walk through occlusion handling (multiple LIDAR points can project to the same pixel — keep the closest). Discuss synchronization (LIDAR and camera timestamps must align).
Explain the difference between a Kalman filter and a particle filter.
Frequently askedOutline
Kalman: optimal for linear systems with Gaussian noise. State + covariance updated via prediction + measurement steps. Extended Kalman for mild non-linearity (linearizes around current estimate). Particle filter: sample-based, handles arbitrary non-linear / non-Gaussian distributions but expensive. Pedestrian motion is non-linear — particle filters or EKF with motion-model selection are common.
How would you handle a perception failure where the model misses a pedestrian in shadow?
Frequently askedOutline
Safety-critical mindset. Discuss redundancy: multiple sensors (LIDAR sees pedestrians regardless of lighting), multi-model ensembles, conservative tracker (if any sensor sees a pedestrian, the system commits to a track). Discuss training-data augmentation (synthetic shadow generation, hard-negative mining). Walk through the 'belt and suspenders' design philosophy in safety-critical perception.
How would you evaluate the quality of a perception model in production?
Frequently askedOutline
Metrics: precision, recall, IoU per class, average precision (AP), MOTA for tracking. Walk through how to compute them on labeled validation data. Discuss closed-loop evaluation (replay logs against the planner) vs open-loop (per-frame metrics). Mention safety case: false negatives on pedestrians are worse than false positives on tree branches.
Waymo interview tips
- C++ AND Python proficiency. Production perception is C++; ML training is Python. Both come up.
- Read 3-5 classic perception papers before your loop: PointNet, Faster R-CNN, CenterPoint, MOTA-tracking literature.
- Sensor calibration, projection, and coordinate frames come up in every perception interview. Know your math cold.
- Safety-critical mindset is the discriminator. The team will probe whether you naturally think about failure modes.
- Compensation per Levels.fyi 2026 carries a specialist premium over generalist SWE. Most candidates have a Master's; PhD bonus for research-oriented roles.
Frequently asked questions
How is Perception Engineer different from generalist SWE at Waymo?
Higher specialist premium, deeper interview on computer vision + ML, more research-oriented day-to-day work. The team owns the perception stack — sensor fusion, detection, tracking, classification.
Do I need a PhD to interview for Perception Engineer?
No, but most candidates have at least a Master's with a vision or robotics specialization. Strong undergrads with research experience also get interviews.
What languages does Waymo's perception stack use?
C++ for production runtime, Python for ML training and tooling, CUDA kernels for GPU-heavy inference paths. Both C++ and Python come up in the interview.
Does Waymo ask machine-learning theory questions for Perception Engineer?
Yes. Expect questions on loss functions, optimization, evaluation metrics, model architectures (Transformers, CNNs, point-cloud networks). One full round is typically ML-focused.
Does Waymo sponsor visas for Perception Engineer new-grad?
Waymo (Alphabet) has sponsored H-1B and OPT in past US cycles. Confirm with your recruiter for 2026.
Practice these live with InterviewChamp.AI
Real-time AI interview assistant that listens to your loop and helps you structure answers under pressure.
Practice these live with InterviewChamp.AI →